Guido Biele
A well described estimand is characterized by
Nobody needs convincing that confounding is a problem
Observed confounders are not a problem, we can simply adjust.
Unobserved confounders are a problem! There is little we can do!
Treatment randomisation prevent confounding.
If the treatment determines who stays in the study and there is another variable L that causes who stays and the outcome, there will be selection bias
Estimating the effect of distance to work out place on fitness will be biased if participation in the follow up (FU) depends on distance, and unobserved education influences both FU & fitness.
An RCT where medication influences side effects (Si), which influence drop out (DO), and unobserved comorbid disorders influence Si & symptoms will produce a biased effect estimate.
Experiments with drop out also require the assumption of no unobserved confounding!
If there are some characteristics C that influence the effect of the treatment T on the outcome O (T-C-interaction), and the distribution of C is different in the study sample and target population, a randomised experiment will produce biased effect estimates.
When the outcomes of interest \(O_1\) and \(O_2\) depend exclusively on a common mediator \(M\) then results about the effect on \(O_1\) will be informative about the effect on \(O_2\)
When only one of the outcomes depends on and additional variable \(X\), results from experiment with outcome \(O_1\) are a biased estimate for the effect on \(O_2\).
Kinge et al., Parental income and mental disorders in children and adolescents: prospective register-based study. IJE (2021).
“…associations between lower parental income and children’s mental disorders were partly, but not fully, attributed to other socio-demographic factors…”
Sariaslan et al., No causal associations between childhood family income and subsequent psychiatric disorders, substance misuse and violent crime arrests: a nationwide Finnish study of >650 000 individuals and their siblings. IJE (2021)
We analyse one data set to see in how far the results of Sariaslan and Kinge depend on modelling assumptions.
When the outcome is binary, sibling designs and panel designs with FE can only use a subset of data: Siblings and individuals with discordant outcomes, respectively.
Choosing a particular study design introduces substantial differences between the study sample and the source and target population.
Association is stronger when parents do not already have a diagnosis.
Original results of Kinge at al (2021) and Sariaslan et al (2021) can be replicated in this data set.
Only a logistic (or cox) regression with adjustment finds an association.